Pith. sign in

REVIEW 3 major objections 5 minor 16 references

An in-silico lung phantom to assess the performance of pulmonary artery segmentation using angiogram

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that synthetic CT lung phantoms with simulated contrast reproduce the pulmonary vasculature to a 2.14% Hausdorff error, providing labeled data for training ML segmentation without contrast injections.

desk verdict The 2.14% Hausdorff error is a self-consistency check of the phantom voxelizer, not evidence that simulated contrast improves real pulmonary artery segmentation; the paper needs a downstream ML or real-image validation before its central claim holds. read the letter →

arxiv 2507.01867 v1 pith:STX4O2TJ submitted 2025-07-02 physics.bio-ph

Sunder Neelakantan , Tanmay Mukherjee , Emilio A. Mendiola , Kyle Myers , Reza Avazmohammadi This is my paper
classification physics.bio-ph
keywords in-silicophantompulmonaryangiographyvasculaturesegmentationsimulatedcontrastHausdorffdistancemachinelearningvalidationfinite-elementmeshhypertension
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that synthetic CT lung images built from a finite-element mesh — in-silico phantoms — can mimic CT pulmonary angiography closely enough to train and validate machine-learning segmentation of the pulmonary vasculature without injected contrast. The authors reconstruct the lungs, central airways, proximal arteries and veins, and distal vessels from a single real CT angiogram, convert that geometry into a mesh, and voxelize the mesh into synthetic images whose vessel intensity can be raised to simulate a contrast bolus. Vasculature segmented from the contrast-simulated phantom agrees with the original mesh to 3.8 mm, or 2.14% of the 140 mm characteristic vessel length. If the approach holds up, it would supply arbitrarily many labeled synthetic images with a built-in ground-truth mesh, supporting contrast-free segmentation and patient-specific modeling of vascular remodeling in pulmonary hypertension.

What carries the argument

The carrying mechanism is proximity-threshold voxelization of the finite-element mesh: mesh nodes are mapped onto a uniform grid, and each voxel's intensity is set by counting how many nodes fall inside a sphere of a chosen radius around it, converting reconstructed geometry into a synthetic CT image. Three such phantoms — lung, proximal pulmonary vessels, and distal vessels with airways — are summed with intensities normalized so that no contrast is apparent, and simulated contrast is produced by weighting the proximal vessel phantom more heavily. The fidelity check is the Hausdorff distance between the mesh segmented from the contrast phantom and the mesh from the original CT image, normalized by the characteristic length of the vasculature.

What would settle it

Train a fixed segmentation model on the simulated-contrast phantoms and evaluate it on held-out real non-contrast and contrast CT scans: if the phantom-trained model does not segment real pulmonary arteries better than an untrained baseline, or if its error on real scans is far above the 2.14% seen in the phantom-to-mesh comparison, the assumption that proximity-voxelized phantoms stand in for real CT imaging is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that simulated contrast — not the geometry alone — is what makes the pulmonary vasculature segmentable. Without contrast, the composite phantom is normalized so that vessels and airways are indistinguishable; with the proximal vessel layer weighted more heavily, the arterial tree stands out and can be reconstructed from the synthetic image. The reconstructed vascular mesh lands within 3.8 mm of the source mesh, a 2.14% error after normalizing by the 140 mm characteristic length, with the remaining difference concentrated in small vessels where 1 mm voxels round off fine detail. The paper presents the method as a controlled platform on which the same phantom can be rendered at multiple simulated contrast levels, with the finite-element mesh doubling as ground truth for machine-learning validation.

Load-bearing premise

The load-bearing premise is that voxelizing a finite-element mesh by proximity counting produces synthetic images that behave like real CT scans for training and validating machine-learning segmentation; the paper's only quantitative test compares the phantom to the very mesh it was built from.

Editorial extensions

If this is right

  • A single reconstructed anatomy can be rendered at many simulated contrast levels, producing large labeled training sets whose ground-truth vasculature is known exactly from the mesh.
  • If contrast-free segmentation becomes reliable, the risk of contrast-agent reactions could be removed from this workflow — the paper cites about 400,000 acute allergic reactions per year tied to contrast-enhanced scans.
  • The vasculature geometry extracted with this method can feed in-silico hemodynamic models, supporting patient-specific studies of pulmonary hypertension remodeling.
  • The paper states that reducing the voxel size below 1 mm should diminish overlap between components and lower the error below 2.14%, at the cost of computation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 2.14% figure measures self-consistency between the phantom and its own source mesh, not fidelity to real CT appearance; comparing phantom images against real non-contrast or low-contrast scans would test the transfer that actually matters.
  • A natural stress-test is to train a small segmentation network on phantoms alone and evaluate it on real non-contrast CT; a sharp performance drop would pinpoint where the proximity-threshold model lacks the noise, blur, and partial-volume effects of real acquisition.
  • The phantom is built from a single representative angiogram and so inherits one patient's anatomy and any segmentation errors in the source image; generating phantoms from many patients would probe generalizability.
  • The binary contrast toggle could be refined into a physiologically parameterized scheme tied to contrast-agent concentration curves, making simulated contrast a tunable variable for robustness testing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper describes a method to generate in-silico CT lung phantoms from a manually segmented finite-element mesh of a single CT pulmonary angiogram from the RSNA Pulmonary Embolism CT Dataset. The phantom is built by mapping mesh nodes onto a uniform grid with a proximity threshold, with an option to boost the intensity of the proximal vasculature layer to simulate contrast. The authors re-segment the proximal vasculature from the contrast-enhanced phantom, compare it to the original mesh via the Hausdorff distance, and report 3.8 mm (2.14% normalized by a characteristic length of 140 mm). From this they conclude that adding simulated contrast can allow successful segmentation of the pulmonary vasculature and that the phantoms can aid ML-based segmentation training and validation.

Significance. If the method were adequately validated, it would provide a useful source of synthetic training data with known ground-truth geometry for ML-based pulmonary artery segmentation, which is a clinically relevant goal given the risks of contrast agents. The pipeline is straightforward, uses an open dataset, and reports a quantitative geometric error metric. However, the reported evidence is a self-consistency check rather than a demonstration of segmentation performance or transfer to real images. No ML segmentation pipeline is exercised, no independent ground truth is used, and the phantom-real image gap is acknowledged but not quantified. The central claim is therefore not yet established, although the approach is plausibly defensible with additional validation.

major comments (3)
  1. [Section II-C, II-D, Eq. (6)] The reported Hausdorff distance of 3.8 mm (2.14%) is a closed-loop self-consistency check: the phantom is generated by voxelizing the same FE mesh that is later used as the reference for comparison. Thresholding an image that contains an explicit intensity boost placed exactly on the known vessel geometry succeeds by construction, so the metric quantifies voxelization and smoothing loss rather than segmentation performance. The paper needs an independent reference, such as comparison against a manually segmented real CT image not used to build the phantom, or at least a repeated-trials analysis to establish uncertainty.
  2. [Abstract, Section III, Section IV] The central claim that 'adding simulated contrast can allow for successful segmentation of the pulmonary vasculature' is not supported by any actual segmentation experiment. The results section only states that increased intensity improves the ability to visualize and segment, and the comparison is between two meshes, not between segmentation outputs. No quantitative segmentation metrics (Dice, precision/recall, or volume overlap) are reported, and no ML model is trained or evaluated. To support the abstract claim, the authors should apply a concrete segmentation method to the phantom with and without simulated contrast and report the resulting performance.
  3. [Section II-C, Section IV] The assumption that proximity-threshold voxelization of an FE mesh behaves like real CT image formation is untested. The paper does not compare the phantom's intensity statistics, noise, partial-volume effects, or point-spread function to real CT data, and the Discussion itself notes that the phantoms 'require further improvement to capture similar levels of detail as the actual images.' Without a quantitative realism check, the transferability of the phantoms to training ML segmentation for real clinical images remains an unsupported premise. A comparison of phantom-derived segmentations against real non-contrast CT images processed by the same method would directly address this gap.
minor comments (5)
  1. [Title page] The running header on the first page identifies the manuscript as 'Physics-informed motion registration of lung parenchyma across static CT images,' which does not match the submitted title or the content of the paper; this should be corrected.
  2. [Section III, first paragraph] There is a duplicated figure caption: 'Fig. 3. Fig. 3. (A) In-silico CT phantom...' The second 'Fig. 3.' should be removed.
  3. [Section II-A, Eq. (1)-(3)] The convolution in Eq. (1) sums over l, m, n from 1 to 3, but the kernels H1 are written as 2D matrices indexed by k; the relationship between the kernel indices and the summation variables should be clarified.
  4. [Section II-D] The characteristic length of 140 mm used to normalize the Hausdorff distance is introduced without explanation; the authors should state how this length is measured from the vasculature mesh.
  5. [Section III, Fig. 3] The paper claims 'varying simulated contrast levels' and 'multiple levels of contrast' but only demonstrates two conditions (with and without contrast). Either show a graded series of contrast levels or temper the wording to match what is shown.

Circularity Check

2 steps flagged · score 6.0 of 10

The 2.14% Hausdorff agreement is a closed-loop check of the phantom voxelizer; the simulated-contrast 'result' is built in by construction.

  1. self definitional [Section II-C/D (phantom generation and comparison)]
    "The in-silico phantom generated with the simulated contrast was used to segment and reconstruct the proximal pulmonary vasculature. Next, the mesh generated through the phantom was compared against the mesh generated from the original image through the use of the Hausdorff distance error metric."

    The mesh 'generated from the original image' is the same manually segmented FE mesh used in Section II-C to create the phantom via voxelization with a proximity threshold. Therefore the comparison is between the source mesh and a re-segmentation of its own voxelized rendering; the reported 3.8 mm / 2.14% error quantifies voxelization and smoothing loss, not the ability to segment new anatomy. The claimed agreement is wired into the input by construction because the phantom is derived from the very mesh used as the reference.

  2. fitted input called prediction [Section II-C and Results (Fig. 3)]
    "Next, the phantoms were added, with the proximal vessel phantom weighted higher to simulate the presence of contrast in the pulmonary vasculature. ... The increased intensity of the proximal vasculature improves the ability to visualize and segment the pulmonary vasculature."

    The simulated contrast is applied precisely to the proximal vessel layer whose geometry was obtained by manual segmentation of the source CT. The reported result that adding this simulated contrast permits successful segmentation is therefore an artifact of construction: thresholding an image whose high-intensity region coincides exactly with the known target must recover that target. No ML segmentation pipeline, no real non-contrast CT, and no noise or partial-volume degradation are tested, so the abstract's conclusion that 'adding simulated contrast can allow for successful segmentation' is not an independent prediction.

full rationale

The paper's central demonstrative result is a self-consistency check: an FE mesh manually segmented from one CT angiogram is voxelized into a phantom, then the phantom is re-segmented and compared to the same source mesh. The 2.14% Hausdorff error therefore measures voxelization and reconstruction fidelity, not the ability of the phantom to represent new anatomy or to transfer to real ML segmentation. The 'contrast improves segmentation' claim is similarly tautological because the contrast is added exactly at the known proximal vessel locations. The Discussion acknowledges these gaps by stating that the phantom generation methods 'require further improvement to capture similar levels of detail as the actual images' and by listing manual segmentation as a limitation, but the abstract still presents the result as established. No external benchmark, real-image test, or ML pipeline is used, so the derivation does not independently support the stated conclusion. Score 6 reflects partial circularity: the reported success is built into the construction, although the paper is candid about its current fidelity limits.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new theoretical entities (particles, forces, or dimensions) are introduced. The in-silico phantom is a simulated object, not a postulated physical entity.

free parameters (6)
  • voxel_size = 1 mm
    Voxel size for phantom generation; manually chosen as comparable to high-resolution CT; directly affects spatial resolution of the phantom.
  • proximity_threshold = 2 mm
    Radius of sphere used to decide if a voxel belongs to a vessel or airway; chosen by hand and not justified in the paper.
  • target_edge_length_lung_mesh = 5 mm
    Target tetrahedral edge length for the lung mesh used in phantom generation.
  • target_edge_length_vessel_mesh = 2 mm
    Target edge length for proximal and distal vessel/airway meshes.
  • contrast_intensity_weight = not specified
    The proximal vessel phantom is 'weighted higher' to simulate contrast, but the exact numerical weight is not given; this determines the contrast level in the synthetic image.
  • hausdorff_normalization_length = 140 mm
    Characteristic length of the vasculature mesh used to normalize the Hausdorff distance; a chosen reference value.
assumptions (4)
  • ad hoc to paper Proximity-threshold voxelization of FE meshes approximates real CT image formation, including contrast behavior.
    This is the core phantom generation step (Section II-C); it maps mesh nodes to voxel intensities via node counts, without modeling CT physics such as attenuation, noise, scatter, or partial volume effects.
  • domain assumption The single RSNA PE CT image selected is representative of clinical CT pulmonary angiography images.
    All reconstructions and phantoms are built from one image (Section II-A); generalizability to other patients and scanners is assumed.
  • domain assumption Manual segmentation in Mimics provides accurate ground truth for the pulmonary vasculature.
    The reference mesh used for comparison and for phantom generation is produced by manual segmentation; any error in this step propagates into the reported Hausdorff distance.
  • domain assumption The Hausdorff distance normalized by a characteristic length is a valid measure of phantom fidelity.
    The paper uses this single metric to conclude the phantom is accurate (Section III), but does not compare it with other segmentation quality metrics or clinical requirements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An in-silico lung phantom to assess the performance of pulmonary artery segmentation using angiogram." pith.science (2026). https://pith.science/paper/STX4O2TJ

@misc{pith2026250701867,
  author       = {Pith},
  title        = {Pith review of: An in-silico lung phantom to assess the performance of pulmonary artery segmentation using angiogram},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STX4O2TJ}},
  note         = {Machine review of arXiv:2507.01867}
}
read the original abstract

Pulmonary hypertension (PH) can lead to significant vascular remodeling, resulting in altered pulmonary blood flow. Estimating the patient-specific contributions of each remodeling event is necessary to optimize and individualize clinical intervention strategies. In-silico modeling has emerged as a powerful tool to simulate pulmonary hemodynamics, and one of the primary requirements for robust in-silico modeling is an accurate representation of the pulmonary vasculature structure. Computed tomography (CT) imaging can be used to segment and reconstruct the proximal vasculature. However, contrast-enhanced imaging, such as CT pulmonary angiography, is required to obtain a comprehensive and high-fidelity view of the pulmonary vasculature. The clinical use of CT pulmonary angiography is limited by the complications associated with the injection of contrast agents. Machine learning (ML) approaches have emerged to effectively segment and reconstruct the pulmonary vasculature without the need for contrast-enhanced imaging. We have developed a method to create in-silico pulmonary angiogram phantoms with varying simulated contrast levels. The results indicated that adding simulated contrast can allow for successful segmentation of the pulmonary vasculature. We expect this method to assist with developing and training ML-based segmentation frameworks and aid in their validation, thereby improving the capability to segment and reconstruct pulmonary vasculature without using contrast-enhanced imaging.

Figures

Figures reproduced from arXiv: 2507.01867 by the authors.

Figure 1
Figure 1. Schematic of the reconstruction and in-silico phantom generation. CT angiogram images were taken from the Radiological Society of North America Pulmonary Embolism CT Dataset [11]. A primary requirement for rigorous in-silico modeling of pulmonary hemodynamics is an accurate representation of the structure of the pulmonary vasculature. Imaging modalities such as computed tomography (CT) and magnetic resonance imaging… view at source ↗
Figure 2
Figure 2. Reconstructed geometry from CT angiogram. (A) Proximal pulmonary arteries and veins. (B) Distal pulmonary vasculature and airways. (C) The left lung with the central airways removed. (D) The final composite mesh used to generate in-silico phantoms. C. In-silico phantom generation The lung geometry was used to generate an in-silico phantom based on our previous work [14]. Briefly, the FE lung mesh nodes were mapped o… view at source ↗
Figure 3
Figure 3. (A) In [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    A global view of pulmonary hypertension,

    M. M. Hoeper, M. Humbert, R. Souza, M. Idrees, S. M. Kawut, K. Sliwa-Hahnle, Z. -C. Jing, and J. S. R. Gibbs, “A global view of pulmonary hypertension,” The Lancet Respiratory Medicine, vol. 4, no. 4, pp. 306–322, 2016

    work page 2016

  2. [2]

    Abstract 14407: Right ventricular myocardial stiffening is associated with pulmonary arterial stiffening in pulmonary hypertension,

    S. Neelakantan, E. P. Manning, P. Zhang, G. Choudhary, and R. Avazmohammadi, “Abstract 14407: Right ventricular myocardial stiffening is associated with pulmonary arterial stiffening in pulmonary hypertension,” Circulation, vol. 148, no. Suppl 1, pp. A14 407–A14 407, 2023

    work page 2023

  3. [3]

    Pulmonary hypertension,

    M. M. Hoeper, H. -A. Ghofrani, E. Gr¨unig et al., “Pulmonary hypertension,” Deutsches Arzteblatt international, vol. 114, no. 5, p. 73— 84, February 2017

    work page 2017

  4. [4]

    Effects of acute rho kinase inhibition on chronic hypoxia -induced changes in proximal and distal pulmonary arterial structure and function,

    R. R. Vanderpool, A. R. Kim, R. Molthen et al., “Effects of acute rho kinase inhibition on chronic hypoxia -induced changes in proximal and distal pulmonary arterial structure and function,” Journal of Applied Physiology, vol. 110, no. 1, pp. 188–198, 2011

    work page 2011

  5. [5]

    Estimating pulmonary arterial pressure difference using integrated machine learning - computational fluid dynamics,

    S. Babak Peighambari, T. Mukherjee, R. Raza Mehdi, E. A. Mendiola, and R. Avazmohammadi, “Estimating pulmonary arterial pressure difference using integrated machine learning - computational fluid dynamics,” in Computing in Cardiology Conference (CinC), vol. 51. Computing in Cardiology, Dec. 2024

    work page 2024

  6. [6]

    Right ventricular architectural remodeling and functional adaptation in pulmonary hypertension,

    E. A. Mendiola, D. da Silva Gonc¸alves Bos, D. M. Leichter, A. Vang, P. Zhang, O. P. Leary, R. J. Gilbert, R. Avazmohammadi, and G. Choudhary, “Right ventricular architectural remodeling and functional adaptation in pulmonary hypertension,” Circulation: Heart Failure, vol. 16, no. 2, p. e009768, 2023

    work page 2023

  7. [7]

    Right ventricular stiffening and anisotropy alterations in pulmonary hypertension: Mechanisms and relations to right heart failure,

    S. Neelakantan, A. Vang, R. R. Mehdi, H. Phelan, P. Nicely, T. Imran, P. Zhang, G. Choudhary, and R. Avazmohammadi, “Right ventricular stiffening and anisotropy alterations in pulmonary hypertension: Mechanisms and relations to right heart failure,” Journal of the American Heart Association, vol. 14, no. 5, p. e037126, 2025

    work page 2025

  8. [8]

    Image -based computational assessment of vascular wall mechanics and hemodynamics in pulmonary arterial hypertension patients,

    B. A. Zambrano, N. A. McLean, X. Zhao et al., “Image -based computational assessment of vascular wall mechanics and hemodynamics in pulmonary arterial hypertension patients,” Journal of Biomechanics, vol. 68, pp. 84–92, 2018

    work page 2018

Show all 16 references
  1. [9]

    Influence of image segmentation on one -dimensional fluid dynamics predictions in the mouse pulmonary arteries,

    M. J. Colebank, L. M. Paun, M. U. Qureshi et al., “Influence of image segmentation on one -dimensional fluid dynamics predictions in the mouse pulmonary arteries,” Journal of The Royal Society Interface, vol. 16, no. 159, p. 20190284, 2019

    2019

  2. [10]

    Dissecting contributions of pulmonary arterial remodeling to right ventricular afterload in pulmonary hypertension,

    S. Neelakantan, E. A. Mendiola, B. Zambrano et al., “Dissecting contributions of pulmonary arterial remodeling to right ventricular afterload in pulmonary hypertension,” Bioeng Transl Med. 2025;e70035

    2025

  3. [11]

    The RSNA pulmonary embolism CT dataset,

    E. Colak, F. C. Kitamura, S. B. Hobbs et al., “The RSNA pulmonary embolism CT dataset,” Radiology: Artificial Intelligence, vol. 3, no. 2, p. e200254, 2021

    2021

  4. [12]

    Automated identification of pulmonary arteries and veins depicted in non -contrast chest CT scans,

    J. Pu, J. K. Leader, J. Sechrist et al., “Automated identification of pulmonary arteries and veins depicted in non -contrast chest CT scans,” Medical Image Analysis, vol. 77, p. 102367, 2022

    2022

  5. [13]

    Physics-informed motion registration of lung parenchyma across static CT images,

    S. Neelakantan, T. Mukherjee, K. Myers, R. Rizi, and R. Avazmohammadi, “Physics-informed motion registration of lung parenchyma across static CT images,” in 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2024, pp. 1–4

    2024

  6. [14]

    In-silico CT lung phantom generated from finite-element mesh,

    S. Neelakantan, T. Mukherjee, B. J. Smith et al., “In-silico CT lung phantom generated from finite-element mesh,” in Proc. of SPIE V ol, vol. 12928, 2024, pp. 1 292 829–1

    2024

  7. [15]

    Berlin, Heidelberg: Springer Berlin Heidelberg, 1998, pp

    Variational Analysis. Berlin, Heidelberg: Springer Berlin Heidelberg, 1998, pp. 108–147

    1998

  8. [16]

    A. C. of Radiology et al., ACR manual on contrast media. American College of Radiology, 2015

    2015

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.